Transcript The Power of Interoperability: Why Objects Are Inevitable Onward! Essay, 2013 http://www.cs.cmu.edu/~aldrich/papers/objects-essay.pdf Comments on this work are welcome.

The Power of Interoperability:
Why Objects Are Inevitable
Onward! Essay, 2013
http://www.cs.cmu.edu/~aldrich/papers/objects-essay.pdf
Comments on this work are welcome. Please send them to aldrich at cmu dot edu
Jonathan Aldrich
Institute for Software Research
School of Computer Science
Carnegie Mellon University
Copyright © 2013 by Jonathan Aldrich. This
work is made available under the terms of the
Creative Commons Attribution-ShareAlike 3.0
license:
http://creativecommons.org/licenses/by-sa/3.0/
Object-Oriented Programming is Widespread
6-8 of top 10 PLs are OO – TIOBE
2
Object-Oriented Programming is Influential
• Major conferences: OOPSLA, ECOOP
• Turing awards for Dahl and Nygaard, and Kay
• Other measures of popularity
•
•
•
•
Langpop.com: 6-8 of most popular languages
SourceForge: Java, C++ most popular
GitHub: JavaScript, Ruby most popular
Significant use of OO design even in procedural languages
• Examples: GTK+, Linux kernel, etc.
• Why this success?
3
OOP Has Been Criticized
“I find OOP technically unsound… philosophically
unsound… [and] methodologically wrong.”
- Alexander Stepanov, developer of the C++ STL
4
Why has OOP been successful?
5
Why has OOP been successful?
“…it was hyped [and] it created a new
software industry.”
- Joe Armstrong, designer of Erlang
Marketing/adoption played a role in the ascent of OOP.
But were there also genuine advantages of OOP?
6
Why has OOP been successful?
“the object-oriented paradigm...is
consistent with the natural way of
human thinking”
- [Schwill, 1994]
OOP may have psychological benefits.
But is there a technical characteristic of
OOP that is critical for modern software?
7
What kind of technical characteristic?
Talk Outline
1. A technical characteristic unique to objects
•
Addressed in Cook’s 2009 Onward! Essay
2. That has a big impact
•
•
Our focus: why that characteristic matters
I.e. how it affects in-the-large software development
8
What Makes OOP Unique?
Candidates: key features of OOP
• Encapsulation?
• Abstract data types (ADTs) also provide
encapsulation
• Inheritance?
• Neither universal nor unique in OOPLs
• Worth studying, but not our focus
• Polymorphism/Dynamic dispatch?
• Every OOPL has dynamic dispatch
• Distinguishes objects from ADTs
animal.speak()
“meow”
“woof”
Dynamic Dispatch as Central to OOP
Significant grounding in the OO literature
• Cook’s 2009 Onward! essay
• Object: “value exporting a procedural interface to data or behavior”
• Objects are self-knowing (autognostic), carrying their own behavior
• Equivalent to Reynolds’ [1975] procedural data structures
• Historical language designs
• “the big idea [of Smalltalk] is messaging” [Kay, 1998 email]
• Design guidance
• “favor object composition over class inheritance” [Gamma et al. ’94]
• “black-box relationships [based on dispatch, not inheritance] are an ideal
towards which a system should evolve” [Johnson & Foote, 1988]
10
Objects vs. ADTs
Two Object-Oriented Sets
interface IntSet {
bool contains(int element)
bool isSubsetOf(IntSet otherSet)
}
Interface is a set of messages
All communication is messagebased; isSubsetOf() implemented by
calling contains() on otherSet
{1}


{1}

{1}
class IntSet1 implements IntSet {…}
class IntSet2 implements IntSet {…}
// in main()
IntSet s1 = new IntSet1(...);
IntSet s2 = new IntSet2(...);
bool x = s1.isSubsetOf(s2);
{1,2}
= true
{1,2}
= true
{1,2}
= true
Set Objects
Different implementations
interoperate freely
11
Objects vs. ADTs
Two Set ADTs
final class IntSetA {
bool contains(int element) { ... }
bool isSubsetOf(IntSetA other) { ... }
}
final class IntSetB {
bool contains(int element) { ... }
bool isSubsetOf(IntSetB other) { ... }
}
// in main()
IntSet sA = new IntSetA(...);
IntSet sB = new IntSetB(...);
bool x = sA.isSubsetOf(sB); // ERROR!
Interface is a set of operations over
a fixed but hidden type (IntSetA)
isSubsetOf() is a binary method that
only works with other instances of
IntSetA. Good for performance.
{1}
{1}
{1}



X
{1,2}
= true
{1,2}
= true
{1,2}
Set ADTs
Different ADT implementations
cannot interoperate
12
Does Interoperability Matter?
• For data structures such as Set, maybe not
• Maybe optimization benefits of ADTs dominate interoperability
“Although a program development support system must store
many implementations of a type..., allowing multiple
implementations within a single program seems less important.”
- A History of CLU [Liskov, 1993]
• But are data structures what OOP is really about?
13
Are Objects “Procedural Data Structures?”
An object is “...a value exporting a procedural interface to data or
behavior.” [Cook, 2009]
“a program execution is regarded as a physical model, simulating the
behavior of either a real or imaginary part of the world”
[Madsen, Møller-Pedersen, Nygaard (and implicitly Dahl), 1993]
“The last thing you wanted any programmer to do is mess with
internal state even if presented figuratively. Instead, the objects
should be presented as sites of higher level behaviors more
appropriate for use as dynamic components.” [Kay, 1993]
14
Service Abstraction
• Objects can implement data structures
• Useful, but not their primary purpose
• Not a unique benefit of objects
• Kay [1993] writes of the “objects as server metaphor” in
which every “object would be a server offering services”
that are accessed via messages to the object.
• A better term is service abstraction
• Definition: a value exporting a procedural interface to behavior
• Identical to procedural data abstraction, but focused on behavior
• Captures the characteristic of objects in which we are interested
15
Service Abstraction provides Interoperability
• Let’s assume service abstraction is the core of OO
• What are the benefits of service abstraction?
• Reynolds/Cook: procedural data abstraction provides
interoperability
• But so do functions, type classes, generic programming, etc.
• What makes service abstraction unique?
16
Interoperability of Widgets
• Consider a Widget-based GUI
• Concept notably developed in Smalltalk
interface Widget {
Dimension getSize();
Dimension getPreferredSize();
void setSize(Dimension size);
void paint(Display display);
Source: http://www.for-a.com/products/hvs300hs/hvs300hs.html
… /* more here */ }
// based on ConstrainedVisual from Apache Pivot UI framework
• Nontrivial abstraction – not just paint()
• A single first-class function is not enough
17
Interoperability of Composite Widgets
• Consider a Composite GUI
• Concept notably developed in Smalltalk
class CompositeWidget implements Widget {
Dimension getSize();
Dimension getPreferredSize();
void setSize(Dimension size);
void paint(Display display);
void add(Widget widget)
… /* more here */ }
// based on Container from Apache Pivot UI framework
Source: http://www.for-a.com/products/hvs300hs/hvs300hs.html
• Nontrivial abstraction – not just paint()
• A single first-class function is not enough
• Composite needs to store diverse subcomponents in a list
• Can’t do this with type classes, generic programming
• Composite needs to invoke paint() uniformly on all subcomponents
• Also breaks type classes, generic programming
18
Design Leverage of Service Abstractions
The ability to define nontrivial abstractions that are modularly
extensible, where instances of those extensions can interoperate in a
first-class way.
• Nontrivial abstractions
• An interface that provides at least two essential services
• Modular extensibility
• New implementations not anticipated when the abstraction was
designed can be provided without changing the original abstraction
• First-class Interoperability
• Interoperability of binary methods
• Such as adding a subcomponent to a composite
• First-class manipulation of different implementations
• Such as putting subcomponents in a list
• Uniform treatment of different implementations
• Such as invoking paint() on all subcomponents
19
Talk Outline
1. A technical characteristic unique to objects
•
•
Objects, for our purposes, are service abstractions that
provide dispatch
Service abstractions uniquely provide first-class
interoperability
2. That has a big impact
•
•
Well, first-class interoperability is nice for GUIs
Does this affect in-the-large software development more
broadly?
20
Large-Scale Development Impact
• How might service abstractions impact in-the-large
software development?
• Some hints
• We are likely looking for an approach to design
• We already know service abstractions are useful for GUIs
• Anecdotally, one can argue that GUIs drove OO
• Smalltalk, MacApp, Microsoft Foundation Classes, Java Applets, …
• What are these GUI designs an instance of?
• A likely candidate: software frameworks [Johnson, 1997]
21
Software Frameworks
• A framework is “the skeleton of an application that can be
customized by an application developer” [Johnson, 1997]
• Frameworks uniquely provide architectural reuse
• Reuse of “the edifice that ties components together”
[Johnson and Foote, 1988]
• Johnson [1997] argues can reduce development effort by 10x
• As a result, frameworks are ubiquitous
•
•
•
•
GUIs: Swing, SWT, .NET, GTK+
Web: Rails, Django, .NET, Servlets, EJB
Mobile: Android, Cocoa
Big data: MapReduce, Hadoop
22
Frameworks need Service Abstraction
• Frameworks define abstractions that extensions implement
• The developer “supplies [the framework] with a set of components that
provide the application specific behavior” [Johnson and Foote, 1988]
• Sometimes the application-specific behavior is just a function
• More often, as we will see, these abstractions are nontrivial
• Frameworks require modular extensibility
• Applications extend the framework without modifying its code
• Frameworks are typically distributed as binaries or bytecode
• cf. Meyer’s [1988] open-closed principle
• Framework developers cannot anticipate the details of extensions
• Though they do plan for certain kinds of extensions
• Frameworks require first-class interoperability
• Plugins often must interoperate with each other
• Frameworks must dynamically and uniformly manage diverse plugins
• We have already seen this for GUI widgets – let’s look at other examples
23
Web Frameworks: Java Servlets
interface Servlet {
void service(Request req, Response res);
void init(ServletConfig config);
void destroy();
String getServletInfo();
ServletConfig getServletConfig();
}
• Nontrivial abstraction
• Lifecycle methods for resource management
• Configuration controls
• Modular extensibility
• Intent is to add new Servlets
• First-class interoperability required
• Web server has a list of diverse Servlet implementations
• Dispatch is required to allow different Servlets to provide their own
behavior
24
Operating Systems: Linux
• Linux is an OO framework!
• In terms of design—not implemented
in an OO language
fat
• File systems as service abstractions
ntfs
• Interface is a struct of function
pointers
• Allows file systems to interoperate
ext2
• E.g. symbolic links between file systems
• Not just file systems
• Many core OS abstractions are extensible
• ~100 Service abstractions in the kernel
25
Operating Systems: Linux
• Linux is an OO framework!
• In terms of design—not implemented
in an OO language
fat
• File systems as service abstractions
ntfs
• Interface is a struct of function
pointers
People
often miss this, or even deny it, but there are manyext2
• Allows file
systems to interoperate
examples
of object-oriented
programming in the kernel.
• E.g. symbolic links between file systems
Although the kernel devel-opers may shun C++ and other
•explicitly
Not justobject-oriented
file systems languages, thinking in terms of
• Manyiscore
OSuseful.
abstractions
are extensible
objects
often
The VFS
[Virtual File System] is a
• ~100
Serviceofabstractions
the kernel
good
example
how to doinclean
and efficient OOP in C,
which is a language that lacks any OOP constructs.
- Robert Love, Linux Kernel Development (2nd Edition)
26
Objection: If I want objects, I can build them!
• Works nicely in a dynamically-typed setting with macros
• Exhibit A: PLT Scheme / Racket
• Works poorly in a statically typed language
• Certainly possible [Kiselyov and Lämmel, 2005]
• Painful in C, Standard ML, Haskell, etc.
• No built-in type gives you exactly what you want
• Annoying object packing/unpacking is necessary
• Feels like an encoding, rather than a natural expression of ideas
• Typed Racket works because of special OO types
• Programmers do it when really necessary
• cf. GTK+ GUI framework, Microsoft COM, Linux drivers, etc.
• My take: people only do this if OO languages are excluded a priori
27
Software Ecosystems
• A software ecosystem is a “set of software solutions that
enable, support, and automate the activities...[of] actors in the
associated social or business ecosystem” [Bosch, 2009]
• Examples: iOS, Android, Windows, Microsoft Office, Eclipse,
Amazon Marketplace, …
• Ecosystems have enormous economic impact
• Driven by network effects [Katz and Shapiro, 1985]
• Top 5 tech firms control or dominate an ecosystem
• Apple, Microsft, IBM, Samsung, Google
• Ecosystems require first-class interoperability
• Critical to achieving benefit from network effects
• “the architecture provides a formalization of the rules of
interoperability and hence teams can, to a large extent, operate
independently” [Bosch, 2009]
28
Mobile Devices: Android
class ContentProvider {
abstract Cursor query(Uri uri, ...);
abstract int insert(Uri uri, ContentValues vals);
abstract Uri update(Uri uri, ContentValues vals, ...);
abstract int delete(Uri uri, ...);
... // other methods not shown
}
• Network effects (apps) give Android value
• Apps build on each other
• Example: contact managers
• Smartr Contacts is a drop-in replacement for the default contact
manager
• Phone, email apps can use Smartr Contacts without preplanning
• Enabled by service abstraction interfaces
• Android keeps a list of heterogeneous ContentProvider
implementations
29
Conclusions
• The essence of objects is dispatch, or service
abstraction
• Dispatch uniquely provides first-class interoperability
• First-class interoperability is critical to frameworks and
ecosystems
• Frameworks and ecosystems are economically critical to
the software industry
30
Hypotheses
• Adding first-class modules to languages without objects
will promote framework-like designs
• Fully parametric module systems will be more practical
with OO types than with ADT types
31
Future Work
• Empirical validation for the benefits of interoperability
• Exploration of other possible benefits of OO
• Psychology
• Inheritance
32
Conclusions
• The essence of objects is dispatch, or service
abstraction
• Dispatch uniquely provides first-class interoperability
• First-class interoperability is critical to frameworks and
ecosystems
• Frameworks and ecosystems are economically critical to
the software industry
33